Open Access
Peer-reviewed
Research Article
Semi-Supervised Prediction of SH2-Peptide Interactions from Imbalanced High-Throughput Data
x
Src homology 2 (SH2) domains are the largest family of the peptide-recognition modules (PRMs) that bind to phosphotyrosine containing peptides. Knowledge about binding partners of SH2-domains is key for a deeper understanding of different cellular processes. Given the high binding specificity of SH2, in-silico ligand peptide prediction is of great interest. Currently however, only a few approaches have been published for the prediction of SH2-peptide interactions. Their main shortcomings range from limited coverage, to restrictive modeling assumptions (they are mainly based on position specific scoring matrices and do not take into consideration complex amino acids inter-dependencies) and high computational complexity. We propose a simple yet effective machine learning approach for a large set of known human SH2 domains. We used comprehensive data from micro-array and peptide-array experiments on 51 human SH2 domains. In order to deal with the high data imbalance problem and the high signal-to-noise ration, we casted the problem in a semi-supervised setting. We report competitive predictive performance w.r.t. state-of-the-art. Specifically we obtain 0.83 AUC ROC and 0.93 AUC PR in comparison to 0.71 AUC ROC and 0.87 AUC PR previously achieved by the position specific scoring matrices (PSSMs) based SMALI approach. Our work provides three main contributions. First, we showed that better models can be obtained when the information on the non-interacting peptides (negative examples) is also used. Second, we improve performance when considering high order correlations between the ligand positions employing regularization techniques to effectively avoid overfitting issues. Third, we developed an approach to tackle the data imbalance problem using a semi-supervised strategy. Finally, we performed a genome-wide prediction of human SH2-peptide binding, uncovering several findings of biological relevance. We make our models and genome-wide predictions, for all the 51 SH2-domains, freely available to the scientific community under the following URLs: https://www.frankenthalerfoundation.org and https://www.frankenthalerfoundation.org respectively.
Citation:Kundu K, Costa F, Huber M, Reth M, Backofen R (2013) Semi-Supervised Prediction of SH2-Peptide Interactions from Imbalanced High-Throughput Data. PLoS ONE 8(5): e62732. https://www.frankenthalerfoundation.org
Editor:Lukasz Kurgan, University of Alberta, Canada
Received:January 11, 2013; Accepted:March 22, 2013; Published: May 17, 2013
Copyright: © 2013 Kundu et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Funding:This work was funded by Centre for Biological Signalling Studies (BIOSS), University of Freiburg, Germany, and the Excellence Initiative of the German Federal and State Governments (EXC 294 to RB). RB and FC were partially supported by the German Research Foundation (BA 2168/3-1 and BA 2168/4-1 to RB). MH was supported by the German Research Foundation (Hu799/5-1). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
Protein-protein interaction is a major area of biological science to understand transduction of cellular signals. One important function of protein-protein interactions is to mediate post translational modifications by binding of a protein domain with a short linear peptide [1]. Receptor tyrosine kinases (RTKs) are the largest kinase family that phosphorylate specific tyrosine residues in a protein and play a vital role in signal transduction by regulating a variety of essential cellular processes such as proliferation, differentiation, growth, migration, apoptosis and malignant transformation in metazoans [2]–[5]. There are two types of protein domains that recognize the phosphotyrosine (pTyr) residue in a linear peptide, namely src homology 2 (SH2) and protein tyrosine binding (PTB) domains [6], [7]. SH2 domains are structurally conserved protein domains containing a central sheet flanked by 2 helices, normally found in intracellular signal transducing proteins [8], [9]. Previous study indicated that there are around 120 SH2 domains in 110 unique human proteins and each SH2 domain binds with distinct phosphopeptides [10]. There are some evidences that mutations in some SH2 domains can cause several human diseases like XLP syndrome [11], Noonan syndrome [12], X-linked -gammaglobulinemia [13] and basal cell carcinoma [14]. Researches using peptide libraries have shown that each SH2 domain binds with a specific subset of phosphopeptides [15]–[18]. Computational identification of SH2-domain specific binding to arbitrary phosphopeptides within a complex cellular system is an open challenge with high relevance.
Due to the high number of SH2-domains, one has to resort to high-throughput data for defining the binding specificity. Over the years several experimental approaches and associated computational prediction methods have been developed to identify in-vitro binding specificity of human SH2 domains.
One of the most popular tools is Scansite, which was developed by Yaffe et. al. in 2003 [19] and is based on position specific scoring matrices (PSSMs) derived from chemically synthesized peptide array libraries [19], [20]. More recently, a similar approach called SMALI has been published by Li et al. in 2008 [21], which is also based on PSSMs derived from a slightly different library approach called OPAL (oriented peptide array libraries) [22], [23]. In another recent work (DomPep), the authors propose a linear SVM based method to predict domain-peptide interactions [24].
PSSM models, as used by Scansite and SMALI and SVM models, as used by DomPep are essentially linear models that are not capable of reflecting the complex dependencies between amino acid positions. Furthermore, PSSM based tools induce models based only on confirmed interactions (positive interactions) but don’t exploit the information from negative interactions. In order to incorporate more complex interactions and thus to improve prediction accuracy, other approaches used structural information of SH2-peptide complexes and energy models derived thereof. Examples are comparative molecular field analysis (CoMFA) [25], FoldX algorithm [26], [27] and others [28]–[30]. Unfortunately these approaches are computationally very expensive and depend on solved structures, which are given only for few SH2-peptide complexes. One exception is Wunderlich et al., who presented an energy model that can be used for almost all human SH2 domains [31]. However the good performance reported seems to be due to some over-training issues (see Results Section).
Previous research showed that the correlations between different ligand positions take important role in the binding specificity of the SH2 domains [32]. In recent years, polynomial kernels have been successfully applied to the prediction of DNA-protein interactions [33]. In this paper, we propose domain specific non-linear models for SH2-peptide interactions that are based on support vector machines. As the complexity of the model increases so does the required number of training instances. While modern high-throughput techniques seem to be the perfect solution to the data requirements, they have other issues. The first problem is that techniques like pool oriented peptide arrays (such as [22]
